Bessemer: capital is a moat, and price matters less

Harry Stebbings opened his conversation with Bessemer's Sameer Dholakia by saying "$100M pre-seed is now normal." He pointed to Bessemer's new funds: $1.75B for early stage and $4B for growth @HarryStebbings. Dholakia's core argument is that markets pick winners and the choice fulfills itself. Capital lets the leader scale fast and build a brand, which then attracts talent and customers. As an example, he says big law firms name just two major legal-AI providers @HarryStebbings.

The deals show what Bessemer will pay:

• Abridge: $700M–$800M valuation at $5M ARR.
• Anthropic: about 150x ARR at the time.
• Legora: led its $1.8B round at $20M ARR.

Dholakia judges each deal on what the company can become in the next 12 months @HarryStebbings.

He also argues venture firms will split into giant global platforms and extreme specialists, and firms in between will struggle to make money @HarryStebbings. He expects companies to stay private longer, with most of this boom's gains captured privately. He still expects IPOs to open up "later next year and in 2028" @HarryStebbings. On models, Bessemer owns both Anthropic and Fireworks so it has exposure to closed and open tokens without having to predict the mix @HarryStebbings.

SimpleClosure's H1 2026 shutdown report gives a different view of the same concentration. It covers only companies SimpleClosure helped dissolve, but it says AI took about 86 cents of every US venture dollar, and $100M+ megadeals took 87.5% of US startup investment. Median age at closure was about 2.3 years for AI companies and 3.8 for SaaS r/SaaS.

Anthropic's data on agents and hiring

SaaStr summarized a report Anthropic published on its own workforce, built from headcount data, agent telemetry, calendars, Slack and hiring scorecards SaaStr. Key numbers:

• Junior hiring: since early 2025, junior hires fell from 27% to 9% of new technical hires. About a third of ML managers and a fifth of software engineering managers said they are hiring less because of AI SaaStr.
• Overall growth: headcount still tripled. Revenue per employee is 2–4x what Google and Meta had at the same age SaaStr.
• Agent hours: by August, agents directed by employees worked 6x as many hours as the employees themselves. Commits per engineer were 5x their level six months earlier SaaStr.
• The bottleneck is people: interactive sessions sit idle about 80% of the time waiting for a human. The authors call the productivity figures directional because adopters were already on rising trajectories SaaStr.
• Hiring criteria: scorecards increasingly reward autonomy and judgment over domain expertise SaaStr.

For founder diligence, the useful point is the 80% idle time: review and orchestration, not generation, is what limits output. A related detail: Anthropic reportedly built its own replacement for git in five months with seven people after merges grew from 2,500 to over 30,000 a day Latent.Space.

AI app founders and the incumbent problem

Former Plaid CTO Jean-Denis Greze shut down an AI tax product for small businesses. It made nearly $1M in its first year and he projected $2.5–3M the next. His reasoning was that incumbents could automate tax prep for their existing customers faster than a startup could win share, and tax is bought once a year @a16z. He pivoted to Town, an email-and-calendar assistant that found product-market fit quickly. Town also competes with OpenAI and Anthropic while buying tokens from them @a16z.

A Disrupt panel described a common path: start on frontier APIs to find product-market fit, then break workflows into tasks and post-train your own models. Fast-growing startups are now switching to open or post-trained models within months of launch TechCrunch. One investor said owning both your models and your harness, not just your data, protects against a provider that later competes with you or raises prices TechCrunch. Another panelist said a meaningful company can now be built with about 10 people TechCrunch.

Neocloud IPOs face pushback

Bloomberg reports Nscale was aiming for about a $30B IPO valuation, while investors think $16–19B would get the deal done. CoreWeave fell 30% over two weeks Bloomberg Podcasts. Microsoft pulled out of a small UK site, and earlier did not become the expected major customer at Nscale's largest site in West Virginia Bloomberg Podcasts. Meanwhile, SemiAnalysis projects +50GW of new capacity in 2027 and $1.3T in leases signed so far this year, and warns the shortage could shift to chips Latent.Space.

Funding and traction

• Greptile (YC W24) raised a $50M Series B. It says it reviews over 1% of GitHub pull requests daily with fewer than 50 people, positioned as an independent check on code regardless of which coding agent wrote it @ycombinator.
• Context.dev (YC S26) raised a $9M seed. It has seven people, handles millions of requests an hour, and has 800+ customers. The founder says he personally takes 300–400 customer and lead calls a month r/SaaS.
• Mercury passed $1B in annualized revenue, has been profitable for 4+ years, and has 450K+ customers. It claims 1 in 3 US startups as customers @immad. It lets external agents act on accounts through MCP, with customer-scoped, revocable permissions and approval required to move money @immad.

Models: open models get cheaper, and policy risk grows

• DeepSeek V4.1 Flash is a 552B-total / 16B-active model under an MIT license. Artificial Analysis ties it for #2 among open-weight models at $0.26 per task, but notes two models score higher at lower cost Latent.Space.
• US ban: a post on X says the US government is banning DeepSeek from civilian federal agency IT systems @MTSlive, and a Bloomberg anchor referenced the same move Bloomberg Podcasts. For startups selling to government, this makes model origin a procurement question.
• Meta's Muse Wonder is priced at $4/$20 per million tokens. Meta's own tests put it level with Claude Fable 5.1 and GPT-6 Astra r/artificial.
• Character.AI filed for Chapter 11. Disney opened the auction with a $20M bid for a non-exclusive license to some of its IP r/artificial.
• Anthropic: reported Q3 revenue of $16.8B ahead of an IPO that could target about $2T r/artificial. Its public-benefit commitment, covered last brief, is structured as a fund that starts at 5% of shares and can grow with valuation. One observer reads it as restricting the share float Latent.Space.

Agent security becomes product

Microsoft expanded Project Perception to continuously assess agent risk, adapt permissions and contain threats. It follows Microsoft's view that models are insider risks whose controls should sit outside the model @satyanadella. Perplexity open-sourced Space, its agent sandbox. Srinivas says building it in-house saved about $15M a year versus a third-party provider, and that the team used frontier models to probe Space for escape routes before release Bloomberg Podcasts.

Deep tech

Quantinuum reports a logical fidelity of five errors per million operations using a color code on its commercial hardware. It says that is 10x better than its own record from last month and 100x better than Google's best published result r/Futurology. Logical errors fell by a factor of 12 with each increase in code distance. Quantinuum says this suggests fault tolerance will need fewer qubits than other approaches r/Futurology.